%load_ext autoreload
%autoreload 2
%matplotlib inline

import numpy as np
import matplotlib
from matplotlib import pyplot as plt
from collections import defaultdict, Counter
import re
import os
import ast
import nltk
import seaborn as sns
import pandas as pd
import pickle

filename_template = '/Users/ff/results_20200303/slurm-{0}.out'
target_template = '/Users/ff/dev/mathematics_dataset-v1.0/{0}-split'
gcloud = False

jobids = ['extrapolate','interpolate']
targetids = jobids

#with open('/Users/ff/dev/mlp-project/scripts/data.pickle', 'rb') as f:
#    lengths_stats = pickle.load(f)
lengths_stats = None

if lengths_stats:
    print(lengths_stats.keys())

extr_tasks = {
    'algebra__polynomial_roots_big_src_test.txt':'algebra__polynomial_roots.txt',
    'arithmetic__add_or_sub_big_src_test.txt':'arithmetic__add_or_sub.txt',
    'arithmetic__add_sub_multiple_longer_src_test.txt':'arithmetic__add_sub_multiple.txt',
    'arithmetic__div_big_src_test.txt':'arithmetic__div.txt',
    'arithmetic__mixed_longer_src_test.txt':'arithmetic__mixed.txt',
    'arithmetic__mul_big_src_test.txt':'arithmetic__mul.txt',
    'arithmetic__mul_div_multiple_longer_src_test.txt':'arithmetic__mul_div_multiple.txt',
    'comparison__closest_more_src_test.txt':'comparison__closest.txt',
    'comparison__kth_biggest_more_src_test.txt':'comparison__kth_biggest.txt',
    'comparison__sort_more_src_test.txt':'comparison__sort.txt',
    'measurement__conversion_src_test.txt':'measurement__conversion.txt',
    'numbers__place_value_big_src_test.txt':'numbers__place_value.txt',
    'numbers__round_number_big_src_test.txt':'numbers__round_number.txt',
    'probability__swr_p_level_set_more_samples_src_test.txt':'probability__swr_p_level_set.txt',
    'probability__swr_p_sequence_more_samples_src_test.txt':'probability__swr_p_sequence.txt'
}

inter_tasks = {
    'algebra__polynomial_roots_src_test.txt':'algebra__polynomial_roots.txt',
    'arithmetic__add_or_sub_src_test.txt':'arithmetic__add_or_sub.txt',
    'arithmetic__add_sub_multiple_src_test.txt':'arithmetic__add_sub_multiple.txt',
    'arithmetic__div_src_test.txt':'arithmetic__div.txt',
    'arithmetic__mixed_src_test.txt':'arithmetic__mixed.txt',
    'arithmetic__mul_div_multiple_src_test.txt':'arithmetic__mul_div_multiple.txt',
    'arithmetic__mul_src_test.txt':'arithmetic__mul.txt',
    'comparison__closest_src_test.txt':'comparison__closest.txt',
    'comparison__kth_biggest_src_test.txt':'comparison__kth_biggest.txt',
    'comparison__sort_src_test.txt':'comparison__sort.txt',
    'measurement__conversion_src_test.txt':'measurement__conversion.txt',
    'numbers__place_value_src_test.txt':'numbers__place_value.txt',
    'numbers__round_number_src_test.txt':'numbers__round_number.txt',
    'probability__swr_p_level_set_src_test.txt':'probability__swr_p_level_set.txt',
    'probability__swr_p_sequence_src_test.txt':'probability__swr_p_sequence.txt'
}

train_names = {'extrapolate':extr_tasks,
               'interpolate':inter_tasks}

#print(train_names['extrapolate']['algebra__polynomial_roots_big_src_test.txt'])

def get_files(lines, regex, end=False, gcloud=True):
    # we have these lines
    # filename
    # [2020-03-03 09:16:02,784 INFO] Translating shard 0.
    # <filename>.txt
    # ....
    # PRED AVG SCORE:
    files = []
    if not end and not gcloud:
        regex = r'Translating shard [0-9]*.'
    for i,line in enumerate(lines):
        match = re.findall(regex, line)
        if match:
            if end:
                filename = lines[i].strip()            
                files.append((filename, i))
            else:
                if gcloud:
                    filename = lines[i].strip()
                    files.append((filename, i))
                else:
                    filename = lines[i-1].strip()
                    files.append((filename, i-1))
    return files

def split_files(raw_lines, filenames_info):
    files = {}
    for i,f in enumerate(filenames_info):
        print(f[0])
        filename = f[0]
        start_line = f[1]
        end_line = f[3]
        
        file_ = raw_lines[start_line:end_line]
        files[i] = (file_)
    return files

class Result:
    def __init__(self, string, prediction, prob, target):
        self.string = string
        self.prediction = prediction
        self.prob = prob
        self.target = target
    
    def __str__(self):
        return (f'String: {self.string}\n\
                 Target: {self.target}\n\
                 Prediction: {self.prediction}\n\
                 Likelihood: {self.prob}')
    
def parse_single_result(string, pred, prob, target):
    string_ = re.sub(r'SENT [0-9]*:\ ', '', string)
    try:
        string_ = ast.literal_eval(string_)
    except ValueError:
        print(string_)
    string = "".join(str(x) for x in string_).replace('_',' ')
    
    pred_ = re.sub(r'PRED [0-9]*:\ ', '', pred)
    pred = "".join(str(x) for x in pred_).strip().replace('_',' ')
    
    prob_ = re.sub(r'PRED SCORE:\ ', '', prob)
    prob = float(prob_)
    
    return Result(string, pred, prob, target)
    
def parse_file_results(raw_lines, targets):
    filename = raw_lines[0].strip()
    _ = raw_lines[1]
    res = [line for line in raw_lines[2:] if line != '\n']
    
    if len(res) % 3 != 0:
        print('Error', filename)
        return
    
    results = []
    for i in range(0, len(res), 3):
        string = res[i]
        prediction = res[i+1]
        prob = res[i+2]
        target = targets[i//3]
        results.append(parse_single_result(string, prediction, prob, target))
    return filename, results

def extract_targets(tasks, targetid):
    targets = {}
    for key in sorted(tasks.keys()):
        task_name = tasks[key][0].strip()
        task_name = task_name.replace('_src_test.txt','')
        path = os.path.join(target_template.format(targetid), task_name+'_tgt_test.txt')
        with open(path, 'r') as f:
            lines = f.readlines()
        lines = [l.strip().replace('_',' ').replace('\n','') for l in lines]
        targets[key] = lines
    return targets

data = {}
for jobid, targetid in zip(jobids,targetids):
    filename = filename_template.format(jobid)
    print(filename)
    
    # Read slurm output
    with open(filename, 'r') as f:
        lines = f.readlines()
    files_start = get_files(lines, r'([a-zA-Z]|_)*.txt\n', gcloud=gcloud)
    files_end = get_files(lines, r'PRED AVG SCORE: [0-9]*.', end=True, gcloud=gcloud)
    
    # Merge these two lists
    merged = sorted([(i[0][0], i[0][1], i[1][0], i[1][1]) for i in zip(files_start,files_end)])
    
    # Split the lines into the different tasks
    tasks = split_files(lines, merged)
    targets = extract_targets(tasks, targetid)
    
    # Create results objects
    data[jobid] = {}
    for key in sorted(tasks.keys()):
        task_name, results = parse_file_results(tasks[key], targets[key])
        data[jobid][task_name] = results

/Users/ff/results_20200303/slurm-extrapolate.out
algebra__polynomial_roots_big_src_test.txt
arithmetic__add_or_sub_big_src_test.txt
arithmetic__add_sub_multiple_longer_src_test.txt
arithmetic__div_big_src_test.txt
arithmetic__mixed_longer_src_test.txt
arithmetic__mul_big_src_test.txt
arithmetic__mul_div_multiple_longer_src_test.txt
comparison__closest_more_src_test.txt
comparison__kth_biggest_more_src_test.txt
comparison__sort_more_src_test.txt
measurement__conversion_src_test.txt
numbers__place_value_big_src_test.txt
numbers__round_number_big_src_test.txt
probability__swr_p_level_set_more_samples_src_test.txt
probability__swr_p_sequence_more_samples_src_test.txt
/Users/ff/results_20200303/slurm-interpolate.out
algebra__polynomial_roots_src_test.txt
arithmetic__add_or_sub_src_test.txt
arithmetic__add_sub_multiple_src_test.txt
arithmetic__div_src_test.txt
arithmetic__mixed_src_test.txt
arithmetic__mul_div_multiple_src_test.txt
arithmetic__mul_src_test.txt
comparison__closest_src_test.txt
comparison__kth_biggest_src_test.txt
comparison__sort_src_test.txt
measurement__conversion_src_test.txt
numbers__place_value_src_test.txt
numbers__round_number_src_test.txt
probability__swr_p_level_set_src_test.txt
probability__swr_p_sequence_src_test.txt

print(data.keys())
print(data['interpolate'].keys())
print(data['extrapolate'].keys())

dict_keys(['extrapolate', 'interpolate'])
dict_keys(['algebra__polynomial_roots_src_test.txt', 'arithmetic__add_or_sub_src_test.txt', 'arithmetic__add_sub_multiple_src_test.txt', 'arithmetic__div_src_test.txt', 'arithmetic__mixed_src_test.txt', 'arithmetic__mul_div_multiple_src_test.txt', 'arithmetic__mul_src_test.txt', 'comparison__closest_src_test.txt', 'comparison__kth_biggest_src_test.txt', 'comparison__sort_src_test.txt', 'measurement__conversion_src_test.txt', 'numbers__place_value_src_test.txt', 'numbers__round_number_src_test.txt', 'probability__swr_p_level_set_src_test.txt', 'probability__swr_p_sequence_src_test.txt'])
dict_keys(['algebra__polynomial_roots_big_src_test.txt', 'arithmetic__add_or_sub_big_src_test.txt', 'arithmetic__add_sub_multiple_longer_src_test.txt', 'arithmetic__div_big_src_test.txt', 'arithmetic__mixed_longer_src_test.txt', 'arithmetic__mul_big_src_test.txt', 'arithmetic__mul_div_multiple_longer_src_test.txt', 'comparison__closest_more_src_test.txt', 'comparison__kth_biggest_more_src_test.txt', 'comparison__sort_more_src_test.txt', 'measurement__conversion_src_test.txt', 'numbers__place_value_big_src_test.txt', 'numbers__round_number_big_src_test.txt', 'probability__swr_p_level_set_more_samples_src_test.txt', 'probability__swr_p_sequence_more_samples_src_test.txt'])

################################
### ACCURACIES
################################


def compute_bleu_corpus_accuracy(sentences):
    hyps = [None] * len(sentences)
    refs = [None] * len(sentences)
    for i, sentence in enumerate(sentences):
        hyps[i] = sentence.prediction
        refs[i] = sentence.target
    corpus_bleu = nltk.translate.bleu_score.corpus_bleu(refs, hyps)
    return corpus_bleu

def compute_bleu_sentence_accuracy(sentences):
    smoothing_function = nltk.translate.bleu_score.SmoothingFunction().method1
    results = np.zeros(len(sentences))
    for i, sentence in enumerate(sentences):
        results[i] = nltk.translate.bleu_score.sentence_bleu(sentence.target, 
                                                             sentence.prediction, 
                                                             smoothing_function=smoothing_function)
    return results

def compute_binary_accuracy(sentences):
    results = np.zeros(len(sentences))
    for i, sentence in enumerate(sentences):
        results[i] = (1 if np.array_equal(sentence.target, sentence.prediction) else 0)
    return results

def compute_nll(sentences):
    results = np.zeros(len(sentences))
    for i, sentence in enumerate(sentences):
        results[i] = sentence.prob
    return results
    
################################
### STATISTICS
################################

def count_numbers_length(sentences):
    float_pattern = "[+-]?(?:[0-9]*[.])?[0-9]+"
    results = np.zeros(len(sentences))
    for i,sentence in enumerate(sentences):
        matches = re.findall(float_pattern, sentence.string)
        lens = [len(match) for match in matches]
        avg_len = sum(lens) / len(lens) if len(lens) > 0 else 0.
        results[i] = avg_len
    return results

def count_number_of_numbers(sentences):
    float_pattern = "[+-]?(?:[0-9]*[.])?[0-9]+"
    results = np.zeros(len(sentences))
    for i,sentence in enumerate(sentences):
        matches = re.findall(float_pattern, sentence.string)
        results[i] = len(matches)
    return results

def ratio_words_numbers(sentences):
    results = np.zeros(len(sentences))
    for i,sentence in enumerate(sentences):
        numbers = sum(c.isdigit() for c in sentence.string)
        words   = sum(c.isalpha() for c in sentence.string)
        results[i] = words / numbers if numbers != 0 else 0
    return results

def len_sentences(sentences):
    results = np.zeros(len(sentences))
    for i,sentence in enumerate(sentences):
        results[i] = len(sentence.string)
    return results

def create_multi_hist(data, n_bins=3, density=False,
                      stats_func = [compute_nll, compute_binary_accuracy, compute_bleu_sentence_accuracy],
                      stats_titles = ['nll', 'binary', 'bleu_sentence'],
                      extra=None):
    
    if extra: n_cols = len(stats_func) + 1
    else: n_cols = len(stats_func)
        
    fig, axs = plt.subplots(nrows=1, ncols=n_cols,  figsize=(10,3))
    fig.tight_layout()
    
    for i,stats in enumerate(stats_func):
        if stats_titles[i] == 'binary':
            width = 0.35
            for bin_ in range(n_bins):
                s = stats(data[bin_])
                axs[i].bar(bin_ - width/2, s.sum(), width, label='1', color='g')
                axs[i].bar(bin_ + width/2, (len(data[bin_]) - s.sum()), width, label='0', color='r')
        else:
            for bin_ in range(n_bins):
                axs[i].hist(stats(data[bin_]), label=bin_, alpha=0.5, density=density)
            axs[i].legend()
        axs[i].set_title(stats_titles[i])
        
    if extra:
        axs[-1].bar(*zip(*extra.items()))
        axs[-1].set_title('train lengths')
        #axs[-1].bar(range(len(extra)), list(extra.values()), align='center')
        #axs[-1].set_xticklabels(range(len(extra)), list(extra.keys()))
        
    plt.show()
        
    #fig, axs = plt.subplots(nrows=1, ncols=1,  figsize=(10,3))
    #fig.tight_layout()
    #for i in range(n_bins):
    #    axs.hist([d.prob for d in data[i]], label=i, alpha=0.5)
    #plt.legend()
    #plt.show()


def compute_statistics(results, function, method='digitize', n_bins=3, density=False,
                       stats_func = [compute_nll, compute_binary_accuracy, compute_bleu_sentence_accuracy],
                       stats_titles = ['nll', 'binary', 'bleu_sentence'],
                       extra=None):
    print(function)
    stats = function(results)

    # Split into bins of fixed width
    if method == 'digitize':
        bins = np.linspace(stats.min(), stats.max(), n_bins, endpoint=False)
        inds = np.digitize(stats, bins)

        results_bins = [[] for i in range(n_bins)]
        for i,v in enumerate(inds):
            results_bins[v-1].append(results[i])

    # Split into bins with fixed number of elements
    # based on sorted results
    elif method == 'equal_split':
        sorted_ = np.argsort(stats)
        results_bins = []
        bins_indices = np.array_split(np.arange(len(results)), n_bins)
        for bin_ in range(n_bins):
            results_bins.append([])
            for i in bins_indices[bin_]:
                results_bins[bin_].append(results[sorted_[i]])        
    else:
        print('WARNING - NO METHOD SELECTED')
        return
        
        
    #for i in range(n_bins):
    #    t = compute_binary_accuracy(results_bins[i])
    #    print(t.mean(), t.std())

    #print(len(results_bins), len(results_bins[0]), len(results_bins[1]), len(results_bins[2]))
    create_multi_hist(results_bins, n_bins=n_bins, density=density, 
                      stats_func=stats_func, stats_titles=stats_titles,
                      extra=extra)

stats_func =  [compute_nll, compute_binary_accuracy]#, compute_bleu_sentence_accuracy]
stats_titles =  ['nll', 'binary']#, 'bleu_sentence']

for targetid in targetids:
    print()
    print('***********************')
    print(targetid)
    print('***********************')
    print()
    for key in sorted(data[targetid].keys()):
        print('-----------------------------------------------')
        print(key)
        results = data[targetid][key]

        task_lengths = {}
        if lengths_stats:
            for d in ['easy','medium','hard']:
                task_lengths[d] = lengths_stats[d][train_names[targetid][key]].mean()
                                
        number_lengths = count_numbers_length(results)
        print('Number lengths: ', number_lengths.mean())

        number_of_numbers = count_number_of_numbers(results)
        print('Number of numbers: ', number_of_numbers.mean())

        words_numbers_ratio = ratio_words_numbers(results)
        print('Words/number ratio: ', words_numbers_ratio.mean())

        length_sentences = len_sentences(results)
        print('Avg. sentence length: ', length_sentences.mean())

        dataset = pd.DataFrame({'Number lengths': number_lengths, 
                                'Number of numbers': number_of_numbers,
                                'Words/number ratio': words_numbers_ratio,
                                'Avg. sentence length': length_sentences})
        
        sns.pairplot(dataset)
        plt.show()

        #binary_acc = compute_binary_accuracy(results)
        #print('Binary: ', binary_acc.mean())

        #print()
        compute_statistics(results, count_numbers_length, method='equal_split', 
                           stats_func=stats_func, stats_titles=stats_titles)
        compute_statistics(results, count_number_of_numbers, method='equal_split', 
                           stats_func=stats_func, stats_titles=stats_titles)
        compute_statistics(results, ratio_words_numbers, method='equal_split', 
                           stats_func=stats_func, stats_titles=stats_titles)
        compute_statistics(results, len_sentences, method='equal_split', 
                           stats_func=stats_func, stats_titles=stats_titles, extra=task_lengths)


        #sentence_acc = compute_bleu_sentence_accuracy(results)
        #print('Sentence: ', sentence_acc.mean())

        #corpus_acc = compute_bleu_sentence_accuracy(results)
        #print('Corpus: ', sentence_acc.mean())

        print('-----------------------------------------------')

***********************
extrapolate
***********************

-----------------------------------------------
algebra__polynomial_roots_big_src_test.txt
Number lengths:  4.619875803012347
Number of numbers:  8.3024
Words/number ratio:  0.4313261691762849
Avg. sentence length:  75.2797

<function count_numbers_length at 0x12817d598>

<function count_number_of_numbers at 0x12817d6a8>

<function ratio_words_numbers at 0x128064620>

<function len_sentences at 0x128523950>

-----------------------------------------------
-----------------------------------------------
arithmetic__add_or_sub_big_src_test.txt
Number lengths:  9.96035
Number of numbers:  2.0
Words/number ratio:  0.5572799690856511
Avg. sentence length:  34.8163

<function count_numbers_length at 0x12817d598>

<function count_number_of_numbers at 0x12817d6a8>

<function ratio_words_numbers at 0x128064620>

<function len_sentences at 0x128523950>